How this course is built

1. Why a video game and not a video

The meta-analysis by Wouters and colleagues (2013) over 77 studies (more than 5,500 participants) finds that "serious games" produce better learning than traditional instruction (d ≈ 0.29) and above all better retention (d ≈ 0.36). But with one decisive detail: they work better when the game is accompanied by other instruction and when the experience is spread over several sessions.

What we take from it here: no level is "just a game". Every mini-game is followed (or preceded) by a written explanation, a formula taken apart and a worked numerical example. And that is why the Flash review exists: to push towards several sessions instead of one marathon.

2. Why you touch before you read

The interactive simulations of the PhET project (University of Colorado Boulder) show conceptual gains equal to or better than those obtained with real lab apparatus, and they produce more productive discussion. The productive failure line of research (Kapur) further shows that attempting a problem before receiving the explanation — even failing — leads to deeper conceptual understanding than being handed the rule up front.

What we take from it here: in almost every level the slider comes before the formula. First you see the arrows cancel, then we tell you it is called destructive interference.

3. Why there are quizzes (and why they do not give grades)

The testing effect (Roediger & Karpicke) is one of the most solid results in cognitive psychology: recalling a piece of information from memory strengthens it far more than re-reading it for the same amount of time. Reviews of study techniques put practice testing and distributed practice at the top for effectiveness.

What we take from it here: the quizzes are low-stakes, with immediate feedback explaining why, endlessly retryable and with no negative marking. Getting it wrong is part of the method: a question you missed comes back sooner in the review, one you got right comes back later (Leitner box system: 1 day → 3 → 7 → 21 → 60).

4. Why levels only unlock through mastery

In mastery learning (Bloom) the student moves on once they have shown mastery of the current unit, with variable time and attempts instead of fixed ones. The measured effects are among the most consistent in individualised instruction, especially for those starting further behind.

5. Why the interface is so bare

Cognitive load theory (Sweller) and the principles of multimedia learning (Mayer) prescribe some very concrete things, applied here to the letter:

6. Colours, contrast, accessibility

• Dark theme with high-contrast text and line width capped at ~75 characters: two parameters that bear directly on reading fatigue.
• No information carried by colour alone: every state also has a symbol or a label (✓, ✗, 🔒), a WCAG requirement for colour-blind readers.
• Touch targets of at least 44 px, visible focus for keyboard navigation, support for prefers-reduced-motion for those who find animation uncomfortable.
• Immediate feedback on every action (under 100 ms): sliders update the drawing as you move them, not on release. That is the threshold beyond which an interface stops feeling like "yours".

7. Motivation: why XP and levels, but no leaderboards

Self-determination theory (Deci & Ryan) identifies three drivers: competence, autonomy, relatedness. Meta-analyses on gamification show positive but fragile effects: points and badges work when they signal competence gained, and can backfire if they become the only reason to play or if they create social-comparison pressure.

What we take from it here: XP arrives only for things genuinely done (mission passed, question recalled), the ranks describe a skill ("Lord of the arrows", "Period hunter") and there is no leaderboard at all, nor a running timer — the reasons, with the research, are in section 22. Autonomy is protected by free mode and by the ability to replay any level.

8. And the last level?

Research on STEM learning distinguishes between being able to apply and being able to transfer. Transfer is only trained on ill-defined problems, where there is no procedure to repeat. That is why the last level is an open workshop: goals to reach, blocks at your disposal, no solution written down anywhere, and a counter of "how many times you queried the oracle" that pushes you to look for a smarter route.

9. Why the qubit arrives at level 1 and Fourier only at 13

The first version of this course followed the "logical" order: waves → Fourier → qubit. Impeccable on paper, wrong in practice: seven levels of maths before seeing the object you came here for.

The current order follows two well-established principles:

• Spiral curriculum (Bruner, 1960): a concept is met more than once, first in simple form and then in full. Amplitudes appear at level 1 as numbers with a sign (+ and −), which is enough to understand interference and H·H; they become complex arrows at level 8, when two signs are no longer enough. Returning to the same concepts after a gap is also distributed practice: two benefits from one choice.
• Just-in-time: a tool is introduced where it is needed, not before. Waves and the Fourier transform arrive at levels 13–17, right before the QFT, once the question "how do I find a hidden periodicity?" has already occurred to the player — at level 12, with Simon's algorithm, you discover you can only find periods "in XOR", and for real ones you need waves. Introducing them seven levels earlier meant answering a question that had not been born yet.

10. Why the classical computer comes first (and stays optional)

"Quantum" is not a thing: it is a difference. And a difference is only visible if you know what it is being compared with. Someone who has never got clear on what a bit is has no way of noticing what is strange about a qubit: they are missing the background the figure stands out against. Nearly every wrong sentence in circulation — "it tries every path at once", "it is infinitely faster" — comes from exactly there.

Hence two choices that hold each other up:

• a Part K of six levels on the normal computer (bits, logic gates, addition with carries, searching, the cost of algorithms, reversibility and Landauer's principle), played with the same mini-games as the rest of the course;
• a comparison block opening every quantum level where a comparison genuinely exists: on the left how it is done on a normal computer, on the right what changes, underneath the number that says how much the difference is worth. Where the comparison is worth a game rather than a paragraph — searching before Grover, the oracle queried by hand before Deutsch–Jozsa, the repetition code before quantum error correction — the classical mini-game is played right there.

Why it is built this way and not as an opening theory chapter:

• Comparing two concrete cases surfaces the principle better than stating the principle and then giving an example (Gentner, Loewenstein & Thompson, 2003; Alfieri, Nokes-Malach & Schunn, 2013). It works especially when the two cases are nearly identical and differ in one point only: which is exactly the relationship between bit and qubit, between linear search and Grover, between DFT and QFT.
• The two columns sit side by side, not one below the other: reading the second while holding the first in mind is cognitive load spent on memory instead of reasoning (the split-attention effect, Sweller, Ayres & Kalyuga). With the columns adjacent, the eyes do the comparing.
• The number at the bottom is always the same kind of number (how many operations, how many queries, how many states): a yardstick that does not change units from one level to the next is what lets you notice that Grover and Shor are not doing the same thing — one is a quadratic gain, the other exponential.
• Part K is optional, like Part 0, and sits outside the prerequisite chain. Anyone who already knows how a computer works walks straight past; anyone with a gap comes back to it the moment the gap starts to hurt — which is the moment you actually learn, not two weeks earlier "because it is on the syllabus". Every comparison block carries a link to the matching classical level, so the door is always open and never a toll gate.

Three details of those four games are deliberate and worth declaring:

• You try first, the explanation comes second, and in classical mode some missions cannot be won. Hitting the wall with the tools you have makes the following solution far more solid than receiving it first: that is Kapur's productive failure, already cited at point 2. The game does say when a wall is a wall, rather than leaving you spinning: failing is useful, failing without knowing it is not.
• The two cases are aligned: if something changes between the modes, it is because it is the thing. That is the condition that makes comparison work (Gentner, structure mapping): the mind isolates the difference only when everything else matches.
• The mode of computation is always visible, on every mini-game in the course: blue "normal computer", violet "quantum computer", with an icon and a written word next to the colour — never colour alone, as at point 6. A pure-maths mini-game — the parabola, al-Khwārizmī's square, the wave labs — carries no badge, because a calculation is neither classical nor quantum: the badge appears only where the game really simulates one of the two machines (the modular clock inside Shor, the quantum gate made of matrices, measurement as a shadow).

11. Sound: information, not decoration

An arcade without sound is not an arcade, but noise for the sake of noise violates Mayer's coherence principle (anything that does not serve the concept steals attention). The compromise applied here:

• each type of event has its own timbre, always the same: click, XP, mission, quantum measurement, constructive interference, destructive interference, level fanfare. After a few minutes you recognise what happened without looking: sound becomes a second information channel;
• short sounds (under 350 ms) at low volume, generated live with WebAudio: no file to download, it works on a slow connection too;
• a mistake does not punish: two descending notes, not a buzzer. Punitive feedback raises performance anxiety and reduces willingness to try again, which is exactly what a mastery-based course needs;
• you can turn it all off and the choice is remembered. No sound starts before an interaction, as browser rules also require.

12. "Wow effect": immediate, specific feedback

In the first version, the Part 0 mini-games did not tell you whether you were doing well. Research on feedback (Hattie & Timperley; Shute) is clear on one point: feedback works when it is immediate, specific and about the task, not the person. Hence three additions, now present in all the games:

• The goal always written at the top, in its own band: you do not have to remember what you were asked.
• A "how close are you" bar going from red to amber to green, with a tone that rises as you approach: colour + sound + text, three channels for the same information (also useful to anyone who struggles with colours).
• Visible celebration at the finish line — sparks, flash, sound — because the moment you understood has to be unmistakable. It is the one "decorative" concession, and it lasts half a second.

13. Why an account is mandatory (and why I do not like it)

Asking for registration is friction, and friction loses people: it would be more convenient for everyone to run everything in the browser. Two reasons make it the right choice anyway:

• progress must not die with the cache. A 28-level course is done across several sessions and often on several devices: losing everything to a browser clean-up is the stupidest way to drop out;
• the certificate has to be worth something. The exam questions come from the server without the correct answers and grading happens on the server: if the state lived in the browser, anyone could award themselves 100% with two lines in the console, and the certificate would be worthless paper.

In exchange: only the data that is genuinely needed (first name, last name, email — date of birth is optional and serves to tell apart people with the same name), no profiling, no advertising, total deletion in one click. It is all written in the privacy notice.

14. The AI tutor that refuses to give you the solution

The tutor answers only from the content of this site (RAG) and, by design, does not provide mission solutions: it gives a hint and points you at the right slider. It is an uncomfortable choice but one supported by the research on desirable difficulties (Bjork): the effort of getting there yourself is precisely what produces durable learning. An obliging tutor would make everything look easier and leave less behind.

Every answer cites the level and links to it: the tutor's goal is to put you back inside the game, not to replace it. And if a question finds no answer in the content, it says so instead of making things up.

15. How do I know the simulator is not lying

A course that teaches with a simulator has a basic problem: if the simulator is wrong, it teaches the error — and teaches it convincingly, because it shows it. The project's tests check the properties you would expect (gates stay unitary, the QFT reproduces the Fourier matrix exactly, probabilities sum to one). But those tests are written by me, on the same reasoning the simulator is written on: if the error is in the reasoning, the tests confirm it instead of catching it.

That is why the simulator is compared against an independent implementation: QuantumSim, written in C by Francesco Sisini. Three hundred randomly generated circuits — up to 4 qubits, with Hadamard, Pauli, S, T, T†, phase rotations, CNOT, CZ and Toffoli — are run through both simulators and the amplitudes compared one by one. Different languages, different authors, code written without knowing each other: the largest gap is of the order of 10⁻¹⁵, that is, the limit of the computer's number precision. An error common to both, at that point, is very unlikely.

QuantumSim is released under the GNU GPL v3 and is not included in this site: it is downloaded and compiled only when the check is run (npm run test:cross), like a bench tool. Thanks to Francesco Sisini for letting me use it — and above all because it is from his books that I started learning this subject.

16. The site's copy: which persuasion techniques I use, declared

This site also has a professional purpose: making me known as someone who builds AI systems and teaches. I think it is only fair to declare which levers I am pulling, so you can weigh them:

17. How it is written for search engines and for AIs

Half of today's searches end up inside an AI-generated answer instead of a list of links. The guidance emerging from research on Generative Engine Optimization is consistent with writing well:

• structured data (JSON-LD: Course, Person, FAQPage) so a machine can tell what this resource is, who wrote it and what it teaches;
• self-contained, quotable statements: every important answer sits in a paragraph that stands on its own, without needing the surrounding context;
• an llms.txt file with the project's verifiable facts, for anyone indexing with language models;
• trust signals: author in the open, linked sources, update date, and explicit admission of the limits (the certificate is not accredited).

Note: the same research shows that the content cited by AIs is the structured and verifiable kind. Another way of saying that writing honestly and writing to be found, for once, coincide.

18. The three languages, and how you move between them

The course exists in full in Italian, English and Spanish: not a translated summary, three complete editions — exam and certificate included. The addresses are translated too (/en/lessons/, /es/lecciones/), because a page in Spanish living in a folder called "lezioni" is a half-translated page, and it shows.

The selector at the top follows four rules, and none of them is about looks:

• Every language is written in its own language — «Español», not «Spanish». People look for their own language the way it is written at home, not translated into a language they may not read. It is the recommendation of the W3C and of the Nielsen Norman Group, and it is also why the first version did not work: it said «IT EN ES», and ISO codes are mute labels to anyone who has never seen them.
• No flags. A flag stands for a state, not a language: which one would you put on Spanish, out of the twenty countries that speak it? In their place a globe, the one symbol the public associates with "language" without associating it with a country.
• You change language without losing your place. Someone reading the QFT in Italian who picks English wants the QFT in English, not to start over from the map. The addresses of the other versions are the same hreflang links the page already declares for search engines: computing them a second time would be the classic way of letting them drift apart.
• If your browser speaks another language I say so, but I do not move you. A single line appears — written in the language it offers, otherwise it cannot be read — with two buttons: switch, or stay. Then it never asks again. Automatic redirection is explicitly discouraged by Google for multilingual sites: it stops you reaching a particular version on purpose, it confuses people who speak more languages than their browser is configured for, and it hides the other copies from search engines.

19. The glossary that stays open while you read

For a while the glossary was the last page of the course. Wrong: the word that blocks your reading is not waiting at the end, you meet it at level 4 — and a word you do not understand does not postpone the problem, it multiplies it, because every sentence after that uses it as if it were clear.

The glossary now sits at the top of every page and stays open while you read. The specific choices come from five research findings, not from a visual fashion:

The panel is deliberately not modal: it does not block the lesson and does not demand to be closed before you can carry on. Blocking the page to show a definition would be like shutting the book to open the dictionary. Below 1100 px of width there is no room for two columns: there it does cover the text, and a tap outside closes it.

A useful side effect: the terms now live in a single file (js/core/glossario.js), so the level 23 table, the panel and the tap-for-definition bubbles can no longer say three different things — which had already started happening between the Italian version and the translations.

20. Funny examples: when a joke teaches and when it gets in the way

Almost everyone carries around some silly thing a teacher said twenty years ago and never forgot. It was put to me like this: «my maths teacher explained the quadratic formula with the dwarfs example, and besides laughing, that stuck with us». The memory is real and the research backs it — but on one precise condition, without which a funny example is not neutral: it makes learning worse, not better.

The result has two halves, and they have to be held together:

• Content-related humour helps. The review by Banas and colleagues (2011) across forty years of studies finds that relevant, non-aggressive, appropriate humour improves attention, climate and willingness to stay on task. The sturdiest effects are on motivation and credibility; on recall the findings are patchier, and the case where it works best is the one where the joke sits inside the explanation and gets picked up again later, rather than being stuck on the side.
• Decorative humour damages. That is the seductive details effect (Harp & Mayer, 1998): interesting but irrelevant details steal attention from the ones that matter. The meta-analyses by Rey (2012) and Sundararajan & Adesope (2020) confirm it: a negative effect on retention and a bigger one still on transfer — that is, precisely on being able to reuse the thing in a new context, which is all that matters here, where middle-school maths has to hold up a quantum level later on.

A third result completes the rule. The keyword method (Pressley, Levin & Delaney, 1982) shows that a vivid mental image which links two things holds them together far better than repetition: that is the same mechanism that makes "the dwarfs" work. But the same line of work dismantles the idea that bizarre is better: what counts is not absurdity, it is the clarity of the image and whether it really ties the two pieces together. A wacky but vague story does nothing; a simple one where you can see who is who stays.

So, in practice, in this course's maths exercises:

• the funny example appears at the point in the reasoning where it is needed, and it carries a step that would otherwise have to be memorised (the dwarfs having to split into two equal rows are the b/2 of completing the square, and the tile left over is the (b/2)²: take them away and the formula goes back to being a nursery rhyme);
• it is picked up again later, in the mini-game and in the quiz, instead of staying an isolated gag — which is the condition under which the humour review finds the best effects on recall;
• there are no images, sounds or anecdotes put there "to lighten things up": those are exactly the seductive details, and it is the same reason why in point 11 the sounds carry information rather than decoration;
• the real history of the mathematics (al-Khwārizmī, Cardano, Hilbert, Schrödinger) is not in the same category: it is not there for laughs, it is there because knowing which problem a formula was born from is part of understanding it — and it lives in a block of its own that you can skip.

An honest note: the humour part is the weakest item on this whole list in terms of strength of evidence — the literature is made of small studies, modest average effects and a lot of variation between teachers. That is why it is used as seasoning on an example that stands up on its own, never as the main vehicle for a concept.

21. Part M: the maths in full, and why it is optional

The course has a rule that holds everywhere: mathematical tools arrive played, right before they are needed, and in the minimum amount needed there. Clock arithmetic arrives before Shor, eigenvectors before phase estimation, the derivative before variational methods. That keeps the course walkable, but it leaves a debt: anyone who wants to know where what they have just used comes from will not find the answer inside the level, because there it would have been dead weight.

Part M pays that debt. It is twelve optional levels covering the syllabus from the first year of secondary school to second-year calculus — but only the pieces this course actually uses, and always with the question «why is it built like that» at the centre: why the quadratic formula has that shape, where the number e comes from, why Grover's √N is a Taylor series stopped at the first term, how you compute e^(iHt), which is how you simulate a molecule.

Why it is organised this way:

• The maths arrives twice, spaced apart. Once stripped to the bone inside the course, and once in full in Part M. That is not wasted repetition: distributed practice is, along with retrieval practice, one of only two techniques Dunlosky and colleagues (2013) rate as high utility in their review of ten study methods. Revisiting the same thing weeks later, in a different form, is worth more than seeing it twice in a row.
• «Why is it built like that» is not an ornament. Asking yourself the reason for what you are learning — elaborative interrogation — and explaining it to yourself in your own words — self-explanation — are two techniques the same review places at moderate utility, and which Chi and colleagues (1994) had shown improve understanding merely by asking students to comment out loud while reading. They are also the two that work best when you already have a foothold in the topic: which is exactly the position of someone arriving at Part M after using those tools in play.
• It is optional, and not out of politeness. The expertise reversal effect (Kalyuga et al., 2003) says the explanation that helps a beginner hinders someone who already knows that piece: they have to read it anyway just to discover they did not need it. Making Part M compulsory would charge a toll to anyone who has already done secondary-school maths. It sits outside the prerequisite chain, and every level of the main course links to its matching M level: the door is open, it is not a gate.
• Every claim is checked by a test, not asserted. When a Part M level says «there are sixteen possible classical strategies and none beats 75%», the project's tests enumerate all sixteen; when it says «the error falls like 1/n», the exponent is measured by doubling the terms and checking what the error gets divided by. It is the same rule the simulator is tested with (section 15), applied to teaching sentences.

Honest note: Part M is the youngest part of the site and the one with the least testing on real users. The choices above are defensible on the literature, but the field evidence — how many start it, how many finish it, whether it actually helps those who do — I do not have yet.

22. Ranks, leaderboards and community

The course shows two numbers, and they are two different things on purpose.

• XP is the score. It goes up when you pass a level and it goes up when you review: every question recalled from the review deck pays five. It has to: answering the same question again days later is distributed practice, which is the reason you will still know something a month from now. A score that only rewards the first time is a score that says «reviewing is wasted time».
• The rank is the progress. It is not a fixed XP threshold but a fraction of the course: half the main path passed is half, whatever the length of the path. Two reasons. The first is measurement: with fixed thresholds, and with XP rising through review, you could reach the top rank by grinding reviews without ever having seen Shor — the rank measured time spent on the site, not what you can do. The second is that the course grows: the old top threshold, set for a course a third of this length, would now fall just past halfway, and the whole second half would stop giving anything. Written as fractions, the thresholds rescale themselves whenever a level is added.

The optional parts pay XP but not ranks, and the levels you do there are counted separately, in plain sight, under the rank line. That is a matter of honesty: a rank says how far you are through the course.

Why there is no leaderboard

It was the first question I asked myself, and the literature answers it more sharply than I expected. The systematic review by Li and colleagues (2024) collects the studies on leaderboards in higher education: they work depending on how they are built, and the most frequently reported problem is a specific one — the discomfort of the learners who see themselves at the bottom of an absolute, public ranking. Which amounts to saying that a leaderboard helps whoever is already ahead and discourages whoever is behind: exactly the person this course exists for.

This is not a theoretical worry. Hanus and Fox (2015) followed two identical courses for a semester, one with a leaderboard and badges and one without: the students in the gamified course declined in intrinsic motivation and satisfaction over the semester, and their lower final exam score was mediated by precisely that decline. And the same leaderboard does not even do the same thing to everyone: a 2024 study in Computers & Education finds that a high ranking mostly helps learners low in trait competitiveness, while a low ranking spurs only those high in it — the same device, on the same page, with opposite signs on different people.

The mechanism was known before gamification existed: when feedback is normative, telling you where you stand against others, whoever does badly attributes it to their own lack of ability and does less on the next task — whereas self-referenced feedback, telling you where you stand against yourself, supports mastery goals, which are the ones tied to interest that lasts (Ames, 1992; on negative feedback and intrinsic motivation, the meta-analysis by Fong and colleagues, 2019).

So instead of a leaderboard there is «Last 7 days», under the rank line: how many levels and how many reviews you did this week, and how many the week before. It is a comparison, but with yourself. It needs no account, no other players, and it has nobody at the bottom.

If a leaderboard ever appears, this is the shape the research supports, and it is worth stating now rather than inventing later: opt-in (you join it, you don't land in it), weekly with a reset (whoever starts today is not chasing three years of head start), in small randomly drawn groups of thirty, showing only your neighbours in the ranking and not the whole table. Never an all-time absolute ranking, and never a yearly one: that is the shape with every flaw described above, multiplied by how long it takes to catch up. There is also a practical reason not to build it now: a leaderboard makes sense when there are enough simultaneous players to fill a group, otherwise it is a list of three names.

Community

Here the research says the opposite, and it is a yes. Relatedness is one of the three needs in self-determination theory, alongside competence and autonomy; and a sense of belonging correlates with mastery goals and with staying (Korpershoek and colleagues, 2020, eighty-two studies). But a community is not a generic chat: the kind that helps learning is tied to the task. The natural place here is level 22 — the workshop where you invent an algorithm of your own: a place to publish yours and look at other people's does exactly the work of comparing cases (see the section on analogies), on top of the work of relatedness.

Honest note: the community does not exist yet, and I will not open one until I can moderate it — an abandoned public space does more harm than none at all. And «Last 7 days» is new: it is defensible on the literature, but on this site I have not yet measured it on real users.

Sources

Listed in the order they appear in the text. Where there is no link it is because the reference is a book or a classic paper that is easy to find: I prefer citing it that way rather than linking a copy of dubious provenance.

• Wouters, P., van Nimwegen, C., van Oostendorp, H., van der Spek, E. (2013). A meta-analysis of the cognitive and motivational effects of serious games. Journal of Educational Psychology, 105(2), 249–265. — ERIC record
• Clark, D. B., Tanner-Smith, E. E., Killingsworth, S. S. (2016). Digital Games, Design, and Learning: A Systematic Review and Meta-Analysis. Review of Educational Research, 86(1), 79–122. — paper
• Sailer, M., Homner, L. (2020). The Gamification of Learning: a Meta-analysis. Educational Psychology Review. — paper
• Perkins, K. et al. PhET: Interactive Simulations for Teaching and Learning Physics. The Physics Teacher, 44(1), 18–23. — paper
• Kapur, M. Productive Failure. — accessible summary (PDF)
• Roediger, H. L., Karpicke, J. D. — on the testing effect and the forward effect of recall: Retrieval practice enhances new learning (PMC)
• Systematic review on spaced learning, interleaving and retrieval practice (2023), Journal of the American College of Radiology. — paper
• Accessible summary on recall + spaced practice: Evidence Based Education
• Bjork, R. A. — Desirable Difficulties in Theory and Practice: why what makes learning harder in the moment makes it more solid over time. — text
• Kivetz, R., Urminsky, O., Zheng, Y. (2006). The Goal-Gradient Hypothesis Resurrected. Journal of Marketing Research. — Columbia Business School summary
• Nunes, J. C., Drèze, X. (2006). The Endowed Progress Effect: How Artificial Advancement Increases Effort. Journal of Consumer Research. — text
• Sisini, F. — QuantumSim, a quantum circuit simulator in C, and the popular books by the same author from which my path into this subject began. — repository
• Gentner, D., Loewenstein, J., Thompson, L. (2003). Learning and Transfer: A General Role for Analogical Encoding. Journal of Educational Psychology — comparing two concrete cases surfaces the shared principle better than studying them one at a time. — PDF
• Alfieri, L., Nokes-Malach, T. J., Schunn, C. D. (2013). Learning Through Case Comparisons: A Meta-Analytic Review. Educational Psychologist, 48(2), 87–113. — article
• Banas, J. A., Dunbar, N., Rodriguez, D., Liu, S.-J. (2011). A Review of Humor in Educational Settings: Four Decades of Research. Communication Education, 60(1), 115–144 — content-related humour helps climate, motivation and attention; aggressive or off-topic humour does not. — text
• Harp, S. F., Mayer, R. E. (1998). How Seductive Details Do Their Damage. Journal of Educational Psychology, 90(3), 414–434 — the interesting but irrelevant detail distracts from what matters.
• Rey, G. D. (2012). A review and a meta-analysis of the seductive detail effect. Educational Research Review, 7(3), 216–237. — article
• Sundararajan, N., Adesope, O. (2020). Keep it Coherent: A Meta-Analysis of the Seductive Details Effect. Educational Psychology Review, 32, 707–734 — negative on retention and, more markedly, on transfer. — article
• Pressley, M., Levin, J. R., Delaney, H. D. (1982). The Mnemonic Keyword Method. Review of Educational Research, 52(1), 61–91 — a mental image that links two things beats repetition; but "bizarre" is not better: what counts is that the image be clear. — article
• Landauer, R. (1961), Irreversibility and Heat Generation in the Computing Process, and Bérut, A. et al. (2012), Experimental verification of Landauer's principle, Nature 483, 187–189 — the thermodynamic cost of erasing a bit, predicted and then measured. — article in Nature
• Bruner, J. S. (1960). The Process of Education — the spiral curriculum: returning to the same concepts at ever deeper levels.
• Sweller, J. — cognitive load theory; Mayer, R. E. — principles of multimedia learning (coherence, signalling, contiguity, segmenting, pre-training, conversational style).
• Hattie, J., Timperley, H. (2007). The Power of Feedback; Shute, V. (2008). Focus on Formative Feedback — feedback works if it is immediate, specific and task-related.
• Deci, E. L., Ryan, R. M. — self-determination theory: competence, autonomy, relatedness.
• Cialdini, R. B. — Influence: reciprocity, social proof, authority. Used here only on verifiable claims.
• On Generative Engine Optimization: overview of practices and position paper on the risks (arXiv)
• WCAG 2.2 (W3C) for contrast, touch targets, visible focus and information never carried by colour alone. — quick reference
• Ayres, P., Sweller, J. — The Split-Attention Principle in Multimedia Learning, in The Cambridge Handbook of Multimedia Learning: why holding a sentence and a definition apart costs working memory.
• Kalyuga, S., Ayres, P., Chandler, P., Sweller, J. (2003). The Expertise Reversal Effect. Educational Psychologist, 38(1), 23–31 — help that serves the beginner hinders the expert.
• Nielsen, J. — 10 Usability Heuristics for User Interface Design, in particular the sixth ("recognition rather than recall"). — text
• Nation, I. S. P. (2001). Learning Vocabulary in Another Language. Cambridge University Press — on the role of glosses.
• Yun, J. (2011). The effects of hypertext glosses on L2 vocabulary acquisition: a meta-analysis. Computer Assisted Language Learning, 24(1), 39–58.
• Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., Willingham, D. T. (2013). Improving Students' Learning With Effective Learning Techniques. Psychological Science in the Public Interest, 14(1), 4–58 — the review that rates ten study techniques: retrieval practice and distributed practice at high utility, elaborative interrogation and self-explanation at moderate utility. — publisher's page
• Chi, M. T. H., de Leeuw, N., Chiu, M.-H., LaVancher, C. (1994). Eliciting self-explanations improves understanding. Cognitive Science, 18(3), 439–477 — explaining out loud what you are reading improves understanding.
• Li, C., Liang, L., Fryer, L. K., Shum, A. (2024). The use of leaderboards in education: A systematic review of empirical evidence in higher education. Journal of Computer Assisted Learning, 40(6), 3406–3442 — the effect depends on the type of leaderboard; the most frequently reported problem is the discomfort of those at the bottom of an absolute, public ranking. — open-access article
• Hanus, M. D., Fox, J. (2015). Assessing the effects of gamification in the classroom: A longitudinal study on intrinsic motivation, social comparison, satisfaction, effort, and academic performance. Computers & Education, 80, 152–161 — a leaderboard and badges for a semester: intrinsic motivation and satisfaction declining, and a lower final score mediated by that decline. — article
• Personalization in educational gamification: Learners with different trait competitiveness benefit differently from rankings on leaderboards (2024). Computers & Education — the same leaderboard has opposite effects on people with different trait competitiveness. — article
• Ames, C. (1992). Classrooms: Goals, Structures, and Student Motivation. Journal of Educational Psychology, 84(3), 261–271 — the difference between a mastery structure (comparison with yourself) and a performance structure (comparison with others), and what the second does to those who are behind.
• Fong, C. J., Patall, E. A., Vasquez, A. C., Stautberg, S. (2019). A Meta-Analysis of Negative Feedback on Intrinsic Motivation. Educational Psychology Review, 31, 121–162. — PDF
• Korpershoek, H., Canrinus, E. T., Fokkens-Bruinsma, M., de Boer, H. (2020). The relationships between school belonging and students' motivational, social-emotional, behavioural, and academic outcomes in secondary education: a meta-analytic review. Research Papers in Education, 35(6), 641–680 — eighty-two studies: belonging correlates with mastery goals and with staying. — article

The whole project is public on GitHub under a free non-commercial licence: studying it, modifying it and using it in state education is always allowed; paid training needs an agreement. If you find an error in these pages, report it.